Crouched over his laptop in his recording studio, he notes the minor harmonies, prominent acoustic guitar, a distant floating organ and delicate background vocals.

He enters all this into a form on a website, and the song is filed away into a vast and ever-growing music database.

Pinkmountain, a musician himself, is one of 30 music analysts working for the world's biggest music streaming service, US-based Pandora. Music streaming allows listeners to listen to music via the internet, without having to download files to their computer. It's like radio, but online.

After nearly 10 years in the role, Pinkmountain has analysed close to 100,000 songs. Fellow analysts include Darian Gray, present drummer for Booker T Jones, and Nate Brenner, bass player from US indie act Tune-Yards. Nate Query of The Decemberists was once an analyst too, but left the company when the band hit global fame.

These are the humans behind the sophisticated mechanics that are changing the way we consume music.

Pandora's analysts map the "DNA" of 10,000 or so songs every month as part of the company's ambitious "Music Genome Project", launched 12 years ago. They record up to 450 unique attributes, such as tone, tempo and instrumentation. From there, complex algorithms match users' listening habits to create a personalised radio experience that serves up songs and artists tailored to their tastes.

Digital technology is redefining our relationship with media, giving consumers more power. It was only a matter of time until this caught up with radio. DJs once decided what you heard, but now that model is breaking down.

Many streaming platforms – including Spotify, Apple's iTunes Radio, Rdio and more – offer some version of personalised music curation.Other offerings include "on-demand" music streaming, which puts the listener in the driving seat, letting them play what they want, when they want. Some platforms pre-program playlists and genre stations; some offer a mix of all the above.

With 77 million active users worldwide at last count, dwarfing its Swedish rival Spotify at 40 million active users, Pandora has trumped them all.

Art and science 

The "magic" of personalised music streaming is a complex science, and each streaming service applies its own formula. Central to each is an attempt to fuse art with science, human with machine; raising the question of whether computers can truly understand something as subjective as music.

The premise is hotly contested. Australian streaming service Guvera rejects the idea completely, insisting that a computer algorithm can never mimic "the human experience associated with discovering new music". Instead, its team of specialists around the globe hand-picks new and diverse playlists daily.

Boston-based Echo Nest, meanwhile, vigorously defends the honour of its algorithms. Acquired by Spotify earlier this year, Echo Nest provides data to music services including Spotify, Rdio and iHeart Radio.

Where Pandora relies on the expert ears of its analysts, Echo Nest uses "machine learning" to catalogue the 3 million-plus songs in its database. Its computer systems learn how best to classify a song by watching how real people do it first.

For instance, explains co-founder Tristan Jehan, Echo Nest asked "hundreds of people" to ascertain the tempo of "thousands of songs".

"From there, the model analyses that data to derive a formula for determining the tempo," he says. "You could expect an algorithm to extract that, but it's actually something very subjective."

The MIT graduate, who boasts a PhD in machine learning, insists there is human input "at every level, in everything we do".

Cultural bias 

Even if one accepts this, it is not without its problems. The people Echo Nest's algorithms learn from are mostly interns from the Boston area, and their musical analysis inevitably comes appended with cultural bias.

Pandora employs experts in genres that may be anomalous to Western ears – such as Sameer Gupta, an Indian classical music teacher and performer.

But Gupta lives in New York City. In fact, a very high proportion of Pandora's music analysts live there, as well as in the San Francisco Bay Area.

In a sense, the mechanical nature of algorithms "abstracts away" from potential cultural bias. But Professor Andrew Brown, who heads up Griffith University's Music Technology department, says algorithms still infer "secondary bias". In other words, the people who design these algorithms bring their background assumptions to that task.

"The assumption that something which is done by computer selection is unbiased is not exactly true," Brown says.

Business interests also tend to skew streaming platforms towards conservatism. They may prioritise adding popular and new artists to their databases, or present obvious recommendations, rather than throwing the listener something left-of-field.

The irony of customised music services resorting to predictability isn't lost on Owen McKern, program manager at Melbourne's PBS FM, a community radio station that has been broadcasting a wide range of specialist music programs since 1979.

McKern says PBS is unlikely to air the same song twice in one week, and believes the beauty of good radio is in the "unexpected".

"You get broadcasters who put songs together that you never would have thought to do in a million years – nor would any computer program," he says. "They just have that X factor to pull completely disparate things out that, for whatever reason, just kinda work."

Rdio, headquartered in San Francisco, goes some way to addressing this: it has a dial for users to "tune" their personalised stations from predictable "favourites" to recommendations that are a little more "adventurous".

Still, it does not fix the problem of streaming services selecting songs largely at random and, in doing so, stripping them of the richness of context.

Professor Brown says there is ample research to show that "telling a story" around a piece of media increases its value – something radio presenters have been doing skilfully for decades. Personalised "radio" services display biographical and historical information about artists, but there's no narrative to contextualise the  delivery of streamed songs.

But given the pace of technological innovation – and the undying faith of its evangelists – it's not much of a leap to imagine how these gaps might be plugged.

Professor Brown thinks it's only a matter of time before streaming services find new ways to enrich the listening experience, but concedes these may not be "human curated".

"It might just be some textual thing that goes along the bottom of the screen, or an or automated voiceover that talks about why the next song is selected – whether or not there's some trajectory from the last three songs."

Stay tuned for personalised voiceovers, coming to an internet radio station near you.